Amazon Patents Humans Assisting Computers

theodp writes "Amazon's latest patent, the Hybrid Machine/Human Computing Arrangement, reads like scary sci-fi, with claims covering the use of humans 'of college educated, at most high school educated, at most elementary school educated, and not formally educated' to perform subtasks dispatched by a computer. From the patent: 'For examples, the task on hand requires French speaking humans, and Task Server has requested that each subtask be performed by at least 10 humans with a past accuracy record of at least 90%.' Yikes."

Comments

This is Amazon's Mechanical Turk system

[Diomidis Spinellis]

Amazon has already deployed such a system under the name of Mechanical Turk. The idea is that humans assist computers, providing what is cutely named artificial artifical intelligence. You can read more about the concept in an article that ACM Queue run on May 2006.
--
Code Quality: The Open Source Perspective

Re:I for one...

[Marxist Hacker 42]

Welcome to Manna. Come to my journal if you want to invest in The Oregon Project, just in case....

Re:So What?

[Marxist Hacker 42]

Just read Marshall Brain's take on the future if a system like Mechanical Turk became the standard for Management in US corporations.

"Human Computation" video

[gaspar ilom]

This video on Human Computation describes using humans as part of a distributed computing grid for interpretting captchas, and categorizing images.

...And they'll actually particpate, en masse -- without pay -- thinking they're just playing an online *game.*

In other news...

[Anonymous Coward]

I played Paranoia and all I got was sued for infringing on Amazon's intellectual property. My character was Ame-R-iCan and his mutant power was eat organic matter x 3.

Already done

[SeaFox]

They already do this at Target.

The employees all wear walkie-talkies and I've heard them come on with an obviously computer synthesized voice telling them a "guest" needed assistance in _____ dept. Or more team members were needed to cashier, ect requesting to know who would address the issue. And they would answer back to it just like they were acknowledging their boss's orders.

Re:Already done

[Manchot]

I think you're looking too much into this. I worked at Target during high school, and what you heard is far less sinister than you believe it to be. The computer telling the "team members" that a "guest" needs assistance is triggered by a customer picking up one of the service phones peppered throughout the store. The reason that the employees try to respond so quickly is that the phones have something written on them to the effect of "A team member will be with you within the next x minutes," and people get angry if you're not there fast enough.

As for the "Additional cashiers to the front lane," that is actually triggered by the cashiers themselves via a button on their registers. Typically, the manager up front would just call for backup directly if there was a sudden wave of customers, but on occasion, the manager would be busy elsewhere. In that case, one of the cashiers could just press the button without stopping (i.e., slowing down) to do it manually. More practically, the cashiers aren't given walkie-talkies because the collective noise from ten nearby walkies would be disruptive to the customers. Also, the phones at each lane were tied into the walkie system, but it didn't work very well, and it was just easier to push the button.

Probably the worst computer-control issue at Target was the "speed score" system. Basically, after every transaction, they'd assign you a score (either G for green, or R for red), indicating whether you were fast enough on that transaction. Your overall scores were then tabulated on a monthly basis. When I first started, you wouldn't know your score on an individual customer, and you could only know your monthly average. Being a 17-year-old, I tried to get the highest score possible, and I did pretty well (something like best average in the store for six months). However, about halfway through my tenure there, they switched to a system that showed you your score after each customer, which soon led me to see how flawed it actually was. You see, I quickly figured out that if not for the customers themselves, I would have gotten a G on every transaction. The problem was that the system basically worked by assigning every item an allowed scan time (so a thing of dog food might be 30 seconds, while a pencil might be 5 seconds). From what I could tell, it also allowed a certain amount of time for payment. What I noticed is that as long as there were a few items (about four or more), I would always get a G, no matter what. When it came down to just a couple of items, I would often get a R. Why? As it turned out, the customer would squander my allowed time by taking a long time to figure out the machine that Target uses for credit and debit cards. In the end, the only effect that speed scores had on me was to get me angry at the people I was supposed to be serving. Yes, you heard that right: I would get a little mad at people for not being efficient enough.

Re:Could an invention like this...

[LilGuy]

That the scary part.

Read manna ( http://www.marshallbrain.com/manna1.htm ) and then see if you'd really want that.

Re:In other news

[grammar fascist]

There's a very good chance this is a semi-supervised machine learning thing.

(Un-disclaimer: I do research in machine learning.)

So you've got this algorithm that, if you give it a bunch of labeled data, it can predict labels for unseen data. (Maybe it labels current best-sellers as likely or unlikely to interest a customer based on his buying habits.) Great. Well, somebody's got to label that data. Human time is expensive. On the other hand, you need as much data as possible: the more the better.

Semi-supervised learning algorithms decide which bits of all their unlabeled data to present to a human for labeling based on how much knowledge it can expect to gain from it. You can get higher accuracy with less data that way.

Here's another problem, though: you don't have that many reliable humans to query. So you make it a game, and pay people based on their accuracy for known things, or on their agreement with humans you've already determined are reliable. Win for the humans: if they're accurate, they get paid. Win for the machine learning algorithm: it gets mounds more accurate, high-information, labeled data. Win for Amazon: they make more sales.

There's nothing scary about this, exactly. It's yet another method of eliciting information from humans that is otherwise very hard or expensive to obtain.